[AutoDiff] Add more Tensor broadcast/unbroadcast differentiation tests.
#24899
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@swift-ci please test tensorflow |
1 similar comment
|
@swift-ci please test tensorflow |
rxwei
added a commit
to rxwei/swift
that referenced
this pull request
May 20, 2019
…rmance. The inefficiency of `unbroadcast(toShape:)`, `unbroadcast(to:)`, and `unbroadcast(like:)` has caused significant performance problems during model training because it's performing a lot of TensorFlow operations to achieve axis calculation. We were forced to implement it this way in the early GPE era when neither send/receive nor per-op dispatch was available. This PR reimplements the unbroadcast operations in terms of host-side logic to compute axes to reduce along. This significantly reduces the TensorFlow opreation dispatch overhead. The base implementation changed from `broadcast(toShape:)` to `broadcast(to:)`. With the new implementation, differentiating broadcasting operators is 37% faster (see simple test script [here](https://gist.github.com/rxwei/e1488cac5379ba2bc3aff7490e18158f)). Note: - Since we now rely on the TensorFlow runtime less, more precondition checks and assertions are added to the newly implemented `unbroadcast(to:)` method. - The part of swiftlang#24408 that uses `Raw.broadcastGradientArgs(s0:s1:)` is still necessary for broadcasting binary operations to become faster. TODO: - Change `unbroadcast(toShape:)` tests added by swiftlang#24899 to use `unbroadcast(to:)`, since `unbroadcast(to:)` is now the base implementation.
rxwei
added a commit
that referenced
this pull request
May 20, 2019
…rmance. (#24907) The inefficiency of `unbroadcast(toShape:)`, `unbroadcast(to:)`, and `unbroadcast(like:)` has caused significant performance problems during model training because it's performing a lot of TensorFlow operations to achieve axis calculation. We were forced to implement it this way in the early GPE era when neither send/receive nor per-op dispatch was available. This PR reimplements the unbroadcast operations in terms of host-side logic to compute axes to reduce along. This significantly reduces the TensorFlow opreation dispatch overhead. The base implementation changed from `broadcast(toShape:)` to `broadcast(to:)`. With the new implementation, differentiating broadcasting operators is 37% faster (see simple test script [here](https://gist.github.com/rxwei/e1488cac5379ba2bc3aff7490e18158f)). Note: - Since we now rely on the TensorFlow runtime less, more precondition checks and assertions are added to the newly implemented `unbroadcast(to:)` method. - The part of #24408 that uses `Raw.broadcastGradientArgs(s0:s1:)` is still necessary for broadcasting binary operations to become faster. TODO: - Change `unbroadcast(toShape:)` tests added by #24899 to use `unbroadcast(to:)`, since `unbroadcast(to:)` is now the base implementation.
|
Closing PR due to refactoring moving |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Similar to the already existing set of tests for
broadcast(toShape:)/unbroadcast(toShape:)in that this adds the same type of tests, but callingbroadcast(to:)/unbroadcast(to:)andbroadcast(like:)/unbroadcast(like:)instead.